Next-Token Prediction

Create a language model based on a body of text and get high-quality predictions (next word, next phrase, next pixel, etc.).

Install

npm i next-token-prediction

Usage

Simple (from a built-in data bootstrap)

Put this /training/ directory in the root of your project.

Now you just need to create your app's index.js file and run it. Your model will start training on the .txt files located in /training/documents/. After training is complete it will run these 4 queries:

const { Language: LM } = require('next-token-prediction');

const MyLanguageModel = async () => {
  const agent = await LM({
    bootstrap: true
  });

  // Predict the next word

  agent.getTokenPrediction('what');

  // Predict the next 5 words

  agent.getTokenSequencePrediction('what is', 5);

  // Complete the phrase

  agent.complete('hopefully');

  // Get a top k sample of completion predictions

  agent.getCompletions('The sun');
};

MyLanguageModel();

---

Advanced (provide trainingData or create it from .txt files)

Put this /training/ directory in the root of your project.

Because training data was committed to this repo, you can optionally skip training, and just use the bootstrapped training data, like this:

const { dirname } = require('path');
const __root = dirname(require.main.filename);

const { Language: LM } = require('next-token-prediction');
const OpenSourceBooksDataset = require(`${__root}/training/datasets/OpenSourceBooks`);

const MyLanguageModel = async () => {
  const agent = await LM({
    dataset: OpenSourceBooksDataset
  });

  // Complete the phrase

  agent.complete('hopefully');
};

MyLanguageModel();

Or, train on your own provided text files:

const { dirname } = require('path');
const __root = dirname(require.main.filename);

const { Language: LM } = require('next-token-prediction');

const MyLanguageModel = () => {
  // The following .txt files should exist in a `/training/documents/`
  // directory in the root of your project

  const agent = await LM({
    files: [
      'marie-antoinette',
      'pride-and-prejudice',
      'to-kill-a-mockingbird',
      'basic-algebra',
      'a-history-of-war',
      'introduction-to-c-programming'
    ]
  });

  // Complete the phrase

  agent.complete('hopefully');
};

MyLanguageModel();

Training embeddings

Note

By default, next-token prediction does not use vector search. To enable it, set VARIANCE=1 (any value higher than 0) in an .env. This will change the prediction from returning the next likeliest token (n-gram search) to returning the most similar token (vector search) e.g. "The quick brown fox jumped..." (n-gram prediction) vs "The quick brown fox juked..." (vector similarity). Note that vector search is considerably slower and more resource intensive.

When running the n-gram training using the built-in training method, vector embeddings (144-dimensional) are also created for each token pair to capture context and semantics (e.g. the token Jordan has different values in the fragment Michael Jordan than it does in the fragment Syria, Jordan). The goal of vector search is to optionally enable paraphrasing, slang and profanity filtering, and more.

Run tests

npm test

Examples

Readline Completion

UI Autocomplete

Videos

readline-completion.mp4

readline-completion-verbose.mp4

Browser example: Fast autocomplete

With more training data you can get more suggestions, eventually hitting a tipping point where it can complete anything.

autocomplete.mp4

Inspiration

3Blue1Brown video on YouTube:

Watch: YouTube

Goals

1. Provide a high-quality text prediction library for:

• autocomplete
• autocorrect
• spell checking
• search/lookup
• summarizing
• paraphrasing

2. Create pixel and audio transformers for other prediction formats

3. Demystify LLMs & simplify methodologies

4. Make a high-quality, free/open chat-focused LLM in JavaScript, and an equally sophisticated image-focused diffusion model. Working on this here.